Re: emacs | Pipeline #8399 has failed for master | ee0e259e

[Philipp Stephani <p.stephani2@gmail.com>]

Am Mo., 4. Jan. 2021 um 18:21 Uhr schrieb Robert Pluim <rpluim@gmail.com>:
>
> Michael Albinus <michael.albinus@gmx.de> writes:
>
> > Hi Robert,
> >
> >>> Is there a way to get src/process-tests.log? That would at least
> >>> narrow it down.
> >>
> >> No, the whole CI job runs into a timeout. I have wrapped now all tests
> >> in process-tests.el with a timeout of 60 seconds; hopefully it shows us
> >> the log file in case of.
> >
> > See <https://emba.gnu.org/emacs/emacs/-/jobs/9636/raw>.
> > process-test-sentinel-sit-for has produced a failure. And according to
> > the timestamps, process-test-stderr-buffer ran into the timeout.
> >
>
> process-test-sentinel-wait-function-working-p which is used by
> process-test-sentinel-sit-for looks racy. Just because the process
> sentinel has been called is no guarantee that the process has exited
> and had its process-status set correctly, perhaps it should loop on
> 'process-live-p'

AFAIK the only correct way to wait for a process to exit is (while
(accept-process-output PROC)).

>
> The process-test-stderr-buffer timeout I can't explain: nothing there
> can wait longer than 2 seconds (unless itʼs 'make-process' thatʼs
> timing out?)

make-process can in theory deadlock on BSD-like systems due to a race
between setting the close-on-exec flag and other fork+exec
combinations. That's rather unlikely though.
But maybe there's a race between accept-process-output and calling the
sentinel so that the sentinel is called too early, and then
accept-process-output blocks? AFAIK, when splitting stderr and stdout,
one has to wait separately for the two involved processes:
(while (accept-process-output process))
(while (accept-process-output stderr-process))